Super powerful! Details of Nvidia Super Chip GB10 Grace Blackwell are revealed

　　The NVIDIA GB10 Grace Blackwell super chip, as the most anticipated product this year, is regarded as its first entry into the desktop CPU market (the mobile terminal is the N1 series), but it was not released as planned in July.

NVIDIA recently released details, emphasizing the results of cooperation with MediaTek, which participated in CPU and interconnect design (previously cooperated in Project DIGITS/DGX Spark mini AI workstation).

　　Technical specifications:

　　· Process and packaging: TSMC 3nm process, 2.5D packaging, CPU (S-dielet) and GPU (G-dielet) separate design　　　· CPU: 20 Armv9.2 cores, divided into two groups (10 cores per group), each group shares 16MB Level 3 cache (independent Level 2 cache undisclosed capacity)　　　　· GPU: Blackwell architecture, the number of CUDA cores is unknown (subjected to be about 6144, equivalent to RTX 5070), equipped with the fifth generation Tensor core, 24MB Level 2 cache, supports ray tracing/DLSS4, FP32 computing power 31 TFLOPS, NVFP4 computing power 1000 TOPS

　· Internet: C2C channel (based on NVLink bus), 16MB system-level cache (can be used as CPU level 4 cache)　· Memory: 256-bit LPDDR5X unified memory, maximum frequency 9400MHz, bandwidth approximately 301GB/s

　· Power consumption: The thermal design power consumption is 140W, and the operating system is a customized DGX Base OS (compatible with other Linux distributions or Windows on Arm unknown)　· Performance: When paired with 128GB of memory, it can run a 200 billion parameter AI model or a 70 billion parameter fine-tuning model; after the two DGX Sparks are interconnected through ConnectX-7, the upper limit of the model parameters has been increased to 405 billion.